API Documentation

sparse_ho:

Models

sparse_ho.models:

ElasticNet([estimator])	Sparse ho ElasticNet model (inner problem).
Lasso([estimator])	Linear Model trained with L1 prior as regularizer (aka the Lasso).
SimplexSVR([estimator])	The simplex support vector regression without bias The optimization problem is solved in the dual.
SparseLogreg([estimator])	Sparse Logistic Regression classifier.
SVM([estimator])	Support Vector Machine classifier without bias.
SVR([estimator])	The support vector regression without bias.
WeightedLasso([estimator])	Linear Model trained with weighted L1 regularizer (aka weighted Lasso).

Criterion

sparse_ho.criterion:

CrossVal(criterion[, cv])	Cross-validation loss.
FiniteDiffMonteCarloSure(sigma[, ...])	Smoothed version of the Stein Unbiased Risk Estimator (SURE).
HeldOutMSE(idx_train, idx_val)	Held out loss for quadratic datafit.
HeldOutSmoothedHinge(idx_train, idx_val)	Smooth Hinge loss.
HeldOutLogistic(idx_train, idx_val)	Logistic loss on held out data
LogisticMulticlass(idx_train, idx_val, algo)	Multiclass logistic loss.

Algorithms

sparse_ho.algo:

Implicit([max_iter, max_iter_lin_sys, ...])	Algorithm to compute the hypergradient using implicit differentiation.
ImplicitForward([tol_jac, max_iter, ...])	Algorithm to compute the hypergradient using implicit forward differentiation.
Forward([use_stop_crit, verbose])	Algorithm to compute the hypergradient using forward differentiation of proximal coordinate descent.
Backward([use_stop_crit, verbose])	Algorithm to compute the hypergradient using backward differentiation.

Optimizers

sparse_ho.optimizers:

Adam([n_outer, epsilon, lr, beta_1, beta_2, ...])	ADAM optimizer for the outer problem.
GradientDescent([n_outer, step_size, ...])	Gradient descent for the outer problem.
LineSearch([n_outer, verbose, ...])	Gradient descent with line search for the outer problem.

Functions

Utils

sparse_ho.utils:

Monitor([callback])

Class used to store computed metrics at each iteration of the outer loop.